The cache a composition's KV scheme names for one full-attention layer
(cache index layer) of a graph whose TurboQuant read is
turboQuantKernels and whose bf16 window read is windowRead (the stock
fused SDPA unless it names another): plain bf16 for bf16 and for a layer a
per-layer scheme does not list, the bf16-first affine lego for affine schemes
(the scheme's read kernels, composedAffineKernels), the bf16-first
TurboQuant lego for turbo. Every cache of one makeCache call shares
masks.
The cache a composition's KV scheme names for one full-attention layer (cache index
layer) of a graph whose TurboQuant read isturboQuantKernelsand whose bf16 window read iswindowRead(the stock fused SDPA unless it names another): plain bf16 forbf16and for a layer a per-layer scheme does not list, the bf16-first affine lego for affine schemes (the scheme's read kernels,composedAffineKernels), the bf16-first TurboQuant lego forturbo. Every cache of onemakeCachecall sharesmasks.